Osmocom.org GTP-U kernel implementation merged mainline

Have you ever used mobile data on your phone or using Tethering?

In packet-switched cellular networks (aka mobile data) from GPRS to EDGE, from UMTS to HSPA and all the way into modern LTE networks, there is a tunneling protocol called GTP (GPRS Tunneling Protocol).

This was the first cellular protocol that involved transport over TCP/IP, as opposed to all the ISDN/E1/T1/FrameRelay world with their weird protocol stacks. So it should have been something super easy to implement on and in Linux, and nobody should have had a reason to run a proprietary GGSN, ever.

However, the cellular telecom world lives in a different universe, and to this day you can be safe to assume that all production GGSNs are proprietary hardware and/or software :(

In 2002, Jens Jakobsen at Mondru AB released the initial version of OpenGGSN, a userspace implementation of this tunneling protocol and the GGSN network element. Development however ceased in 2005, and we at the Osmocom project thus adopted OpenGGSN maintenance in 2016.

Having a userspace implementation of any tunneling protocol of course only works for relatively low bandwidth, due to the scheduling and memory-copying overhead between kernel, userspace, and kernel again.

So OpenGGSN might have been useful for early GPRS networks where the maximum data rate per subscriber is in the hundreds of kilobits, but it certainly is not possible for any real operator, particularly not at today's data rates.

That's why for decades, all commonly used IP tunneling protocols have been implemented inside the Linux kernel, which has some tunneling infrastructure used with tunnels like IP-IP, SIT, GRE, PPTP, L2TP and others.

But then again, the cellular world lives in a universe where Free and Open Source Software didn't exit until OpenBTS and OpenBSC changed all o that from 2008 onwards. So nobody ever bothered to add GTP support to the in-kernel tunneling framework.

In 2012, I started an in-kernel implementation of GTP-U (the user plane with actual user IP data) as part of my work at sysmocom. My former netfilter colleague and current netfilter core team leader Pablo Neira was contracted to bring it further along, but unfortunately the customer project funding the effort was discontinued, and we didn't have time to complete it.

Luckily, in 2015 Andreas Schultz of Travelping came around and has forward-ported the old code to a more modern kernel, fixed the numerous bugs and started to test and use it. He also kept pushing Pablo and me for review and submission, thanks for that!

Finally, in May 2016, the code was merged into the mainline kernel, and now every upcoming version of the Linux kernel will have a fast and efficient in-kernel implementation of GTP-U. It is configured via netlink from userspace, where you are expected to run a corresponding daemon for the control plane, such as either OpenGGSN, or the new GGSN + PDN-GW implementation in Erlang called erGW.

You can find the kernel code at drivers/net/gtp.c, and the userspace netlink library code (libgtpnl) at git.osmocom.org.

I haven't done actual benchmarking of the performance that you can get on modern x86 hardware with this, but I would expect it to be the same of what you can also get from other similar in-kernel tunneling implementations.

Now that the cellular industry has failed for decades to realize how easy and little effort would have been needed to have a fast and inexpensive GGSN around, let's see if now that other people did it for them, there will be some adoption.

If you're interested in testing or running a GGSN or PDN-GW and become an early adopter, feel free to reach out to Andreas, Pablo and/or me. The osmocom-net-gprs mailing list might be a good way to discuss further development and/or testing.